Submission SBERT_Large_mt_ru_finetuning

May 29, 2021, 6:54 a.m.

Team: SberDevices

Model url: https://huggingface.co/sberbank-ai/sbert_large_mt_nlu_ru


Total score: 0.514

Dataset Score Metric
LiDiRus 0.218 Matthew`s Corr
RCB 0.351 / 0.486 F1/Acc
PARus 0.498 Accuracy
MuSeRC 0.642 / 0.319 F1a/Em
TERRa 0.637 Accuracy
RUSSE 0.657 Accuracy
RWSD 0.675 Accuracy
DaNetQA 0.697 Accuracy
RuCoS 0.35 / 0.347 F1/EM
Model description:

BERT large model multitask (cased) for Sentence Embeddings in Russian language. The model was fine-tuned on NER, toxic, Dialogue and RussianSuperGLUE tasks, code written on tensorflow. For better quality we use for sentence representation mean token embeddings.


Parameter description:

Diagnostic (Matthew`s Correlation): 0.218

Category Score
LOGIC 0.16192611130344906
KNOWLEDGE 0.22740828237325783
PREDICATE-ARGUMENT STRUCTURE 0.20139781205785764
LEXICAL SEMANTICS 0.2813434604687791
Lexical Semantics - Lexical Entailment 0.2148928384990026
Lexical Semantics - Morphological Negation 0.2828894749305018
Lexical Semantics - Factivity 0.1781741612749496
Lexical Semantics - Symmetry/Collectivity 0.22821773229381923
Lexical Semantics - Redundancy 0.30692490702242564
Lexical Semantics - Named Entities 0.16692446522239718
Lexical Semantics - Quantifiers 0.47387910220727386
Predicate-Argument Structure Core Args 0.3304857994026801
Predicate-Argument Structure Prepositional Phrases 0.26354595321588503
Predicate-Argument Structure Ellipsis/Implicits 0.23570226039551584
Predicate-Argument Structure Anaphora/Coreference -0.002651969513727742
Predicate-Argument Structure Active/Passive 0.1382147381437882
Predicate-Argument Structure Nominalization 0.24552983583446694
Predicate-Argument Structure Genitives/Partitives 0.10482848367219183
Predicate-Argument Structure Datives 0.3563483225498992
Predicate-Argument Structure Relative Clauses 0.01642880193633814
Predicate-Argument Structure Coordination Scopes 0.17910620335162064
Predicate-Argument Structure Intersectivity 0.4129955231527934
Predicate-Argument Structure Restrictivity -0.1895424836601307
Logic Negation 0.09642916726065977
Logic Double Negation 0.2727272727272727
Logic Interval/Numbers -0.019157088122605363
Logic Conjuction 0.10540925533894599
Logic Disjunction 0.1286978904175574
Logic Conditionals 0.14285714285714285
Logic Universal 0.26856632724128343
Logic Existential 0.060522753266880246
Logic Temporal 0.12434118282549846
Logic Upward Monotone 0.3398363161942986
Logic Downward Monotone 0.10984700727621793
Logic Non-Monotonic -0.2163856337073152
Knowledge Common Sense 0.2514207350602513
Knowledge World Knowledge 0.1854598701600781

Performance:

Dataset Speed RAM
LiDiRus - -
RCB - -
PARus - -
MuSeRC - -
TERRa - -
RUSSE - -
RWSD - -
DaNetQA - -
RuCoS - -